scroll the issue ↓
Robot learning from video, VLAs, world models, JEPA — overnight we rank ~1,600 items from arXiv, X, Reddit, Hacker News, the trade press and Substack down to the ~30 that matter to the field. A real sample issue is right below. Scroll it.
Six firehoses, ranked overnight, cut to the ~30 items worth your morning.
This is the actual pipeline behind the sample above — same stages, every night. The per-source counts below are from one real run.
Each night's full arXiv announcement batch, plus the ranked top of everything else.
Every item is ranked for topical relevance — no editors asleep at the wheel, no pay-to-place. The keywords are printed in the issue's footer, so you can audit what the ranking optimizes for.
The survivors are laid out as a three-page broadsheet — front page, research, newsletters, the discourse. Titles and summaries stay in the source's own words; every item keeps its link.
One PDF, five minutes, done — the exact artifact you scrolled at the top of this page. Founding readers get the very first delivered issues.
Why we show you this: a $5 digest earns trust with its process, not its adjectives. Slow news days make thinner issues — the pipeline never pads.
I'm Siddhesh Kanawade. I built the pipeline behind Pixels2Actions because I was drowning in a dozen tabs every morning trying to keep up with robot learning. The selection is algorithmic; the standards are mine — I'll read every issue before it ships, and when you reply to one, a human answers.
Researchers and engineers working on robot learning, imitation learning, VLAs and world models — and the founders and investors tracking physical AI. If you open arXiv, X and Hacker News in three tabs before breakfast, this triages all three for you.
The selection is algorithmic: items are ranked for topical relevance, and each issue's footer lists the keywords its curation targeted — scroll to page 3 of the sample above and you can read them. The words are not generated: titles and summaries are the source's own — a paper's abstract, an article's lede — and every item links to the original. Nothing is paraphrased by a model.
They're weekly and cover all of robotics; we're daily and deliberately narrow — robot learning from video, teleoperation and demonstration data, VLAs, and world models (JEPA and beyond). If you want breadth, read them (we do too — they feed our newsletter section). If this is your field, read us.
A PDF is finite — front page, an end, no infinite scroll. Read it offline, print it, archive it. It's a morning paper, because that's the right shape for a morning read.
Narrow is the point — broad robotics digests already exist, but the video-to-policy and world-models firehose has no daily brief. Every reader gets the same three pages today; the ranking runs on a keyword profile, so editions for other niches are the natural next build, and founding subscribers get them first.
Reply to any issue or email us and you're out, effective immediately — we refund the current month on request, no questions. A one-click cancel link ships with the first issues.
Tomorrow's issue gets built whether you're on the list or not.